#Deep Q Network for ENO
#Resetting the battery to BOPT on each day during training
#Increase no. of actions to 10

%matplotlib inline

import pulp
import matplotlib.pyplot as plt
from matplotlib.colors import ListedColormap

import pandas as pd
import numpy as np
from random import shuffle
from mpl_toolkits.mplot3d import Axes3D

import torch
import torch.nn as nn
import torch.nn.functional as F

np.random.seed(230228)

# Hyper Parameters
BATCH_SIZE = 24
LR = 0.01                   # learning rate
EPSILON = 0.9               # greedy policy
GAMMA = 0.9                 # reward discount
LAMBDA = 0.8                # parameter decay
TARGET_REPLACE_ITER = 24*7*4*2    # target update frequency (every month)
MEMORY_CAPACITY = 24*7*4*6      # store upto four month worth of memory   

N_ACTIONS = 10 #no. of duty cycles (0,1,2,3,4)
N_STATES = 4 #number of state space parameter [batt, enp, henergy, fcast]
HIDDEN_LAYER = 30

class ENO(object):
    def __init__(self, year=2010):
        self.year = year
        self.day = None
        self.hr = None
        
        self.TIME_STEPS = None
        self.NO_OF_DAYS = None
        
        self.BMIN = 0.0
        self.BMAX = 20000.0 #Battery capacity
        self.BOPT = 0.6 * self.BMAX #Assuming 60% of battery is the optimal level
        self.HMAX = 1000
        
        self.senergy = None #matrix with harvested energy data for the entire year
        self.fforecast = None #matrix with forecast values for each day
        
        self.batt = None #battery variable
        self.enp = None #enp at end of hr
        self.henergy = None #harvested energy variable
        self.fcast = None #forecast variable
    
    #function to map total day energy into day_state
    def get_day_state(self,tot_day_energy):
        if (tot_day_energy < 2500):
            day_state = 0
        elif (2500 <= tot_day_energy < 5000):
            day_state = 1
        elif (5000 <= tot_day_energy < 8000):
            day_state = 2
        elif (8000 <= tot_day_energy < 10000):
            day_state = 3
        elif (10000 <= tot_day_energy < 12000):
            day_state = 4
        else:
            day_state = 5
        return int(day_state)

    #function to get the solar data for the given year and prep it
    def get_data(self):
        filename = str(self.year)+'.csv'
        #skiprows=4 to remove unnecessary title texts
        #usecols=4 to read only the Global Solar Radiation (GSR) values
        solar_radiation = pd.read_csv(filename, skiprows=4, encoding='shift_jisx0213', usecols=[4])
        
        #convert dataframe to numpy array
        solar_radiation = solar_radiation.values
        solar_energy = np.array([i *0.0165*1000000*0.15*1000/(60*60) for i in solar_radiation])
        
        #reshape solar_energy into no_of_daysx24 array
        _senergy = solar_energy.reshape(-1,24)
        _senergy[np.isnan(_senergy)] = 0 #convert missing data in CSV files to zero
        self.senergy = _senergy
        
        
        #create a perfect forecaster
        tot_day_energy = np.sum(_senergy, axis=1) #contains total energy harvested on each day
        get_day_state = np.vectorize(self.get_day_state)
        self.fforecast = get_day_state(tot_day_energy)
        
        return 0
    
    def reset(self):
        
        self.get_data() #first get data for the given year
        
        self.TIME_STEPS = self.senergy.shape[1]
        self.NO_OF_DAYS = self.senergy.shape[0]
        
        print("Environment is RESET")
        
        self.day = 0
        self.hr = 0
        
        self.batt = self.BOPT #battery returns to optimal level
        self.enp = self.BOPT - self.batt #enp is reset to zero
        self.henergy = self.senergy[self.day][self.hr] 
        self.fcast = self.fforecast[self.day]
        
        state = [self.batt/self.BMAX, self.enp/(self.BMAX/2), self.henergy/self.HMAX, self.fcast/5] #normalizing all state values within [0,1] interval
        reward = 0
        done = False
        info = "RESET"
        return [state, reward, done, info]
    
    
    #reward function
    def rewardfn(self):
        mu = 0
        sig = 1000
#         return ((1./(np.sqrt(2.*np.pi)*sig)*np.exp(-np.power((self.enp - mu)/sig, 2.)/2)) * 2000000)-400


        if(np.abs(self.enp) <= 2400): #24hr * 100mW/hr
            return ((1./(np.sqrt(2.*np.pi)*sig)*np.exp(-np.power((self.enp - mu)/sig, 2.)/2)) * 1000000)
        else:
            return -100 - 0.05*np.abs(self.enp)
    
    def step(self, action):
        done = False
        info = "OK"
#         print("Next STEP")
        
        reward = 0
        e_consumed = (action+1)*50
        
        self.batt += (self.henergy - e_consumed)
        self.batt = np.clip(self.batt, self.BMIN, self.BMAX)
        self.enp = self.BOPT - self.batt
        
        if(self.hr < self.TIME_STEPS - 1):
            self.hr += 1
            self.henergy = self.senergy[self.day][self.hr] 
        else:
            if(self.day < self.NO_OF_DAYS -1):
                reward = self.rewardfn() #give reward only at the end of the day
                self.hr = 0
                self.day += 1
                self.henergy = self.senergy[self.day][self.hr] 
                self.fcast = self.fforecast[self.day]
            else:
                reward = self.rewardfn()
                done = True
                info = "End of the year"
                
        _state = [self.batt/self.BMAX, self.enp/(self.BMAX/2), self.henergy/self.HMAX, self.fcast/5]
        return [_state, reward, done, info]

class Net(nn.Module):
    def __init__(self, ):
        super(Net, self).__init__()
        self.fc1 = nn.Linear(N_STATES, HIDDEN_LAYER)
        self.fc1.weight.data.normal_(0, 0.1)   # initialization
        
        self.fc2 = nn.Linear(HIDDEN_LAYER, HIDDEN_LAYER)
        self.fc2.weight.data.normal_(0, 0.1)   # initialization
        
        self.fc3 = nn.Linear(HIDDEN_LAYER, HIDDEN_LAYER)
        self.fc3.weight.data.normal_(0, 0.1)   # initialization
        
        self.fc4 = nn.Linear(HIDDEN_LAYER, HIDDEN_LAYER)
        self.fc4.weight.data.normal_(0, 0.1)   # initialization
        
        self.out = nn.Linear(HIDDEN_LAYER, N_ACTIONS)
        self.out.weight.data.normal_(0, 0.1)   # initialization

    def forward(self, x):
        x = self.fc1(x)
        x = F.relu(x)
        actions_value = self.out(x)
        return actions_value

class DQN(object):
    def __init__(self):
        self.eval_net, self.target_net = Net(), Net()
        print("Neural net")
        print(self.eval_net)

        self.learn_step_counter = 0                                     # for target updating
        self.memory_counter = 0                                         # for storing memory
        self.memory = np.zeros((MEMORY_CAPACITY, N_STATES * 2 + 2))     # initialize memory [mem: ([s], a, r, [s_]) ]
        self.optimizer = torch.optim.Adam(self.eval_net.parameters(), lr=LR)
        self.loss_func = nn.MSELoss()

    def choose_action(self, x):
        x = torch.unsqueeze(torch.FloatTensor(x), 0)
        # input only one sample
        if np.random.uniform() < EPSILON:   # greedy
            actions_value = self.eval_net.forward(x)
            action = torch.max(actions_value, 1)[1].data.numpy()
            action = action[0] # return the argmax index
        else:   # random
            action = np.random.randint(0, N_ACTIONS)
            action = action
        return action
    
    def choose_greedy_action(self, x):
        x = torch.unsqueeze(torch.FloatTensor(x), 0)
        # input only one sample
    
        actions_value = self.eval_net.forward(x)
        action = torch.max(actions_value, 1)[1].data.numpy()
        action = action[0] # return the argmax index

        return action

    def store_transition(self, s, a, r, s_):
        transition = np.hstack((s, [a, r], s_))
        # replace the old memory with new memory
        index = self.memory_counter % MEMORY_CAPACITY
        self.memory[index, :] = transition
        self.memory_counter += 1

    def learn(self):
        # target parameter update
        if self.learn_step_counter % TARGET_REPLACE_ITER == 0:
            self.target_net.load_state_dict(self.eval_net.state_dict())
        self.learn_step_counter += 1

        # sample batch transitions
        sample_index = np.random.choice(MEMORY_CAPACITY, BATCH_SIZE)
        b_memory = self.memory[sample_index, :]
        b_s = torch.FloatTensor(b_memory[:, :N_STATES])
        b_a = torch.LongTensor(b_memory[:, N_STATES:N_STATES+1].astype(int))
        b_r = torch.FloatTensor(b_memory[:, N_STATES+1:N_STATES+2])
        b_s_ = torch.FloatTensor(b_memory[:, -N_STATES:])

        # q_eval w.r.t the action in experience
        q_eval = self.eval_net(b_s).gather(1, b_a)  # shape (batch, 1)
        q_next = self.target_net(b_s_).detach()     # detach from graph, don't backpropagate
        q_target = b_r + GAMMA * q_next.max(1)[0].view(BATCH_SIZE, 1)   # shape (batch, 1)
        loss = self.loss_func(q_eval, q_target)

        self.optimizer.zero_grad()
        loss.backward()
        self.optimizer.step()

dqn = DQN()
eno = ENO(2010)
NO_OF_ITERATIONS = 100
avg_reward_rec = np.empty(1)
for iteration in range(NO_OF_ITERATIONS):
    print('\nCollecting experience... Iteration:', iteration)
    s, r, done, info = eno.reset()
    record = np.empty(4)

    while True:
    #     print([eno.day, eno.hr])

        a = dqn.choose_action(s)
    #     print("Actin is ",a)
        #state = [batt, enp, henergy, fcast]
        record = np.vstack((record, [s[0],s[2],r, a])) #record battery, henergy, reward and action
    #     print("Action is" , a)
        # take action
        s_, r, done, info = eno.step(a)
    #     print([s_,r])
    #     print("\n")
        if eno.hr == 0:
            eno.batt = eno.BOPT #resetting the battery to the optimal value for each day
        dqn.store_transition(s, a, r, s_)

        if dqn.memory_counter > MEMORY_CAPACITY:
            dqn.learn()

        if done:
            print("End of Data")
            break

        s = s_

    record = np.delete(record, 0, 0) #remove the first row which is garbage

    reward_rec = record[:,2]
    reward_rec = reward_rec[reward_rec != 0]
    print("Average reward =", np.mean(reward_rec) )
    avg_reward_rec = np.append(avg_reward_rec, np.mean(reward_rec))

    action_rec = record[:,3]

    fig = plt.figure(figsize=(10,5))

    ax1 = fig.add_subplot(1,2,1)
    ax1.plot(reward_rec,'y')
    plt.ylabel("REWARD")
    plt.xlabel("Day")
    ax1.set_ylim([-400,400])

    ax2 = fig.add_subplot(1,2,2)
    plt.hist(action_rec, rwidth=0.75)#     plt.ylabel("Action")

    fig.tight_layout()
    plt.show()

avg_reward_rec = np.delete(avg_reward_rec, 0, 0) #remove the first row which is garbage
plt.plot(avg_reward_rec,'b')

Neural net
Net(
  (fc1): Linear(in_features=4, out_features=30, bias=True)
  (fc2): Linear(in_features=30, out_features=30, bias=True)
  (fc3): Linear(in_features=30, out_features=30, bias=True)
  (fc4): Linear(in_features=30, out_features=30, bias=True)
  (out): Linear(in_features=30, out_features=10, bias=True)
)

Collecting experience... Iteration: 0
Environment is RESET
End of Data
Average reward = -160.24184385598102

Collecting experience... Iteration: 1
Environment is RESET
End of Data
Average reward = -75.22163881527506

Collecting experience... Iteration: 2
Environment is RESET
End of Data
Average reward = -63.559301985629546

Collecting experience... Iteration: 3
Environment is RESET
End of Data
Average reward = -39.114646534755806

Collecting experience... Iteration: 4
Environment is RESET
End of Data
Average reward = -63.16003673144816

Collecting experience... Iteration: 5
Environment is RESET
End of Data
Average reward = -16.784645615285566

Collecting experience... Iteration: 6
Environment is RESET
End of Data
Average reward = -5.245425754490298

Collecting experience... Iteration: 7
Environment is RESET
End of Data
Average reward = 29.239301601860408

Collecting experience... Iteration: 8
Environment is RESET
End of Data
Average reward = 14.312884613529999

Collecting experience... Iteration: 9
Environment is RESET
End of Data
Average reward = 5.293823247919756

Collecting experience... Iteration: 10
Environment is RESET
End of Data
Average reward = 16.604781327977758

Collecting experience... Iteration: 11
Environment is RESET
End of Data
Average reward = 17.828528745108358

Collecting experience... Iteration: 12
Environment is RESET
End of Data
Average reward = 7.125340418205079

Collecting experience... Iteration: 13
Environment is RESET
End of Data
Average reward = 8.085487831875549

Collecting experience... Iteration: 14
Environment is RESET
End of Data
Average reward = 11.081115453341479

Collecting experience... Iteration: 15
Environment is RESET
End of Data
Average reward = 42.646987466949994

Collecting experience... Iteration: 16
Environment is RESET
End of Data
Average reward = 44.68982926374964

Collecting experience... Iteration: 17
Environment is RESET
End of Data
Average reward = 46.11494774501377

Collecting experience... Iteration: 18
Environment is RESET
End of Data
Average reward = 44.49212549727419

Collecting experience... Iteration: 19
Environment is RESET
End of Data
Average reward = 49.4383819118609

Collecting experience... Iteration: 20
Environment is RESET
End of Data
Average reward = 51.4539586205768

Collecting experience... Iteration: 21
Environment is RESET
End of Data
Average reward = 42.894173368770964

Collecting experience... Iteration: 22
Environment is RESET
End of Data
Average reward = 49.79876952494771

Collecting experience... Iteration: 23
Environment is RESET
End of Data
Average reward = 50.352718150738305

Collecting experience... Iteration: 24
Environment is RESET
End of Data
Average reward = 41.63279763478672

Collecting experience... Iteration: 25
Environment is RESET
End of Data
Average reward = 46.10338097854738

Collecting experience... Iteration: 26
Environment is RESET
End of Data
Average reward = 15.656679668786268

Collecting experience... Iteration: 27
Environment is RESET
End of Data
Average reward = 26.588875844510714

Collecting experience... Iteration: 28
Environment is RESET
End of Data
Average reward = -1.0588654944198783

Collecting experience... Iteration: 29
Environment is RESET
End of Data
Average reward = -17.449747041373566

Collecting experience... Iteration: 30
Environment is RESET
End of Data
Average reward = 8.401195636228515

Collecting experience... Iteration: 31
Environment is RESET
End of Data
Average reward = -6.911316304983682

Collecting experience... Iteration: 32
Environment is RESET
End of Data
Average reward = 8.260830382523876

Collecting experience... Iteration: 33
Environment is RESET
End of Data
Average reward = 29.10912953112777

Collecting experience... Iteration: 34
Environment is RESET
End of Data
Average reward = -17.099342903260904

Collecting experience... Iteration: 35
Environment is RESET
End of Data
Average reward = -49.27286899943259

Collecting experience... Iteration: 36
Environment is RESET
End of Data
Average reward = 2.1671541227033457

Collecting experience... Iteration: 37
Environment is RESET
End of Data
Average reward = 12.63746038407406

Collecting experience... Iteration: 38
Environment is RESET
End of Data
Average reward = -2.6903928966184423

Collecting experience... Iteration: 39
Environment is RESET
End of Data
Average reward = 3.5647614118425577

Collecting experience... Iteration: 40
Environment is RESET
End of Data
Average reward = -3.134298798881326

Collecting experience... Iteration: 41
Environment is RESET
End of Data
Average reward = 21.14828809557825

Collecting experience... Iteration: 42
Environment is RESET
End of Data
Average reward = 58.92770745064865

Collecting experience... Iteration: 43
Environment is RESET
End of Data
Average reward = 33.99214661827062

Collecting experience... Iteration: 44
Environment is RESET
End of Data
Average reward = 46.04116277996911

Collecting experience... Iteration: 45
Environment is RESET
End of Data
Average reward = 41.70466729480486

Collecting experience... Iteration: 46
Environment is RESET
End of Data
Average reward = 24.55066925597246

Collecting experience... Iteration: 47
Environment is RESET
End of Data
Average reward = 7.932485120098535

Collecting experience... Iteration: 48
Environment is RESET
End of Data
Average reward = -25.362063209492

Collecting experience... Iteration: 49
Environment is RESET
End of Data
Average reward = 4.622508519843375

Collecting experience... Iteration: 50
Environment is RESET
End of Data
Average reward = 13.019074303373666

Collecting experience... Iteration: 51
Environment is RESET
End of Data
Average reward = 31.375758199042508

Collecting experience... Iteration: 52
Environment is RESET
End of Data
Average reward = -4.331727456886716

Collecting experience... Iteration: 53
Environment is RESET
End of Data
Average reward = 25.56523423782647

Collecting experience... Iteration: 54
Environment is RESET
End of Data
Average reward = 18.346671391154548

Collecting experience... Iteration: 55
Environment is RESET
End of Data
Average reward = 33.23916122456389

Collecting experience... Iteration: 56
Environment is RESET
End of Data
Average reward = 30.496220554161933

Collecting experience... Iteration: 57
Environment is RESET
End of Data
Average reward = 38.30772491536196

Collecting experience... Iteration: 58
Environment is RESET
End of Data
Average reward = 19.84038969491561

Collecting experience... Iteration: 59
Environment is RESET
End of Data
Average reward = 9.614662932415795

Collecting experience... Iteration: 60
Environment is RESET
End of Data
Average reward = 15.36850191417485

Collecting experience... Iteration: 61
Environment is RESET
End of Data
Average reward = 0.9301612508573941

Collecting experience... Iteration: 62
Environment is RESET
End of Data
Average reward = -7.747550774790257

Collecting experience... Iteration: 63
Environment is RESET
End of Data
Average reward = 2.358196060592713

Collecting experience... Iteration: 64
Environment is RESET
End of Data
Average reward = -2.198265658234474

Collecting experience... Iteration: 65
Environment is RESET
End of Data
Average reward = 25.43478769554863

Collecting experience... Iteration: 66
Environment is RESET
End of Data
Average reward = -8.973692050602205

Collecting experience... Iteration: 67
Environment is RESET
End of Data
Average reward = -0.2366998716771667

Collecting experience... Iteration: 68
Environment is RESET
End of Data
Average reward = 15.290083997621624

Collecting experience... Iteration: 69
Environment is RESET
End of Data
Average reward = 2.2539282782773316

Collecting experience... Iteration: 70
Environment is RESET
End of Data
Average reward = -1.5764095480954583

Collecting experience... Iteration: 71
Environment is RESET
End of Data
Average reward = 1.7559613227854443

Collecting experience... Iteration: 72
Environment is RESET
End of Data
Average reward = -15.039993894545974

Collecting experience... Iteration: 73
Environment is RESET
End of Data
Average reward = -3.2608955593522055

Collecting experience... Iteration: 74
Environment is RESET
End of Data
Average reward = 15.353368683695493

Collecting experience... Iteration: 75
Environment is RESET
End of Data
Average reward = 3.682650344376765

Collecting experience... Iteration: 76
Environment is RESET
End of Data
Average reward = 14.417269356662619

Collecting experience... Iteration: 77
Environment is RESET
End of Data
Average reward = 4.350543027088681

Collecting experience... Iteration: 78
Environment is RESET
End of Data
Average reward = -9.534355131175543

Collecting experience... Iteration: 79
Environment is RESET
End of Data
Average reward = 6.197487985999532

Collecting experience... Iteration: 80
Environment is RESET
End of Data
Average reward = 26.659185540958127

Collecting experience... Iteration: 81
Environment is RESET
End of Data
Average reward = 24.743348289465292

Collecting experience... Iteration: 82
Environment is RESET
End of Data
Average reward = 5.231494029236323

Collecting experience... Iteration: 83
Environment is RESET
End of Data
Average reward = 9.114212741071778

Collecting experience... Iteration: 84
Environment is RESET
End of Data
Average reward = 8.443260528495792

Collecting experience... Iteration: 85
Environment is RESET
End of Data
Average reward = -8.711691869627195

Collecting experience... Iteration: 86
Environment is RESET
End of Data
Average reward = 27.97483647838269

Collecting experience... Iteration: 87
Environment is RESET
End of Data
Average reward = 15.51204524760664

Collecting experience... Iteration: 88
Environment is RESET
End of Data
Average reward = 13.84166538904309

Collecting experience... Iteration: 89
Environment is RESET
End of Data
Average reward = -16.936603032691217

Collecting experience... Iteration: 90
Environment is RESET
End of Data
Average reward = -15.288494451828887

Collecting experience... Iteration: 91
Environment is RESET
End of Data
Average reward = -3.2651184612752213

Collecting experience... Iteration: 92
Environment is RESET
End of Data
Average reward = -25.256227002612693

Collecting experience... Iteration: 93
Environment is RESET
End of Data
Average reward = -3.717648013291267

Collecting experience... Iteration: 94
Environment is RESET
End of Data
Average reward = -37.025844345152684

Collecting experience... Iteration: 95
Environment is RESET
End of Data
Average reward = -42.87862076514412

Collecting experience... Iteration: 96
Environment is RESET
End of Data
Average reward = -10.428381174359027

Collecting experience... Iteration: 97
Environment is RESET
End of Data
Average reward = -56.425877845007925

Collecting experience... Iteration: 98
Environment is RESET
End of Data
Average reward = 31.047804827709435

Collecting experience... Iteration: 99
Environment is RESET
End of Data
Average reward = 24.11861646167388

[<matplotlib.lines.Line2D at 0x7fe6070b3cc0>]

print('\nTesting...')
s, r, done, info = eno.reset()
test_record = np.empty(4)

while True:
#     print([eno.day, eno.hr])

    a = dqn.choose_greedy_action(s)
    
    #state = [batt, enp, henergy, fcast]
    test_record = np.vstack((test_record, [s[0],s[2],r, a])) #record battery, henergy, reward and action
#     print("Action is" , a)
    # take action
    s_, r, done, info = eno.step(a)
#     print([s_,r])
#     print("\n")
    if eno.hr == 0:
        eno.batt = eno.BOPT #resetting the battery to the optimal value for each day
   
    if done:
        print("End of Data")
        break
       
    s = s_

Testing...
Environment is RESET
End of Data

test_reward_rec = test_record[:,2]
test_reward_rec = test_reward_rec[test_reward_rec != 0]
plt.plot(test_reward_rec)

[<matplotlib.lines.Line2D at 0x7fe607007a90>]

plt.plot(test_record[:,0],'r')

[<matplotlib.lines.Line2D at 0x7fe602cd5278>]

#Average Battery Percentage
np.mean(test_record[:,0])

0.6820840229425865

for DAY in range(eno.NO_OF_DAYS):
    START = DAY*24
    END = START+24

    fig = plt.figure(figsize=(10,4))
    st = fig.suptitle("DAY %s" %(DAY))

    ax1 = fig.add_subplot(141)
    ax1.plot(test_record[START:END,0])
    ax1.set_title("Battery")
    ax1.set_ylim([0,1])

    ax2 = fig.add_subplot(142)
    ax2.plot(test_record[START:END,1])
    ax2.set_title("Harvested Energy")
    ax2.set_ylim([0,1])

    ax3 = fig.add_subplot(144)
    ax3.axis('off')
    if END < (eno.NO_OF_DAYS*eno.TIME_STEPS):
        plt.text(0.5, 0.5, "REWARD = %.2f\n" %(test_record[END+1,2]),fontsize=14, ha='center')

    ax4 = fig.add_subplot(143)
    ax4.plot(test_record[START:END,3])
    ax4.set_title("Action")
    ax4.set_ylim([0,N_ACTIONS])

    fig.tight_layout()
    st.set_y(0.95)
    fig.subplots_adjust(top=0.75)
    plt.show()